Micro light vision assisted enhancement system and method for glaucoma patients

Abstract

The application discloses a micro-light vision auxiliary enhancement system and method for glaucoma patients and relates to the technical field of image processing. The system comprises: a data acquisition module, which acquires an environment image, a bias parameter and a weight coefficient; a metric extraction module, which performs multi-scale band decomposition on the environment image, calculates the energy ratio of an edge region, generates a perceived gap degree in combination with the bias parameter, obtains signal saliency by using a noise amplitude, and obtains cross-scale stability based on the structural stability between adjacent frequency bands; a decision coupling module, which takes the perceived gap degree as a gate, constructs a correction compensation term by using the weight coefficient, the stability and the signal saliency complement, and thus obtains a calibrated saliency; and an image reconstruction module, which constructs a mapping relationship between the above parameters to modulate each sub-band coefficient, and obtains an enhanced image through inverse transform processing. The application realizes image frequency domain adaptive modulation based on a modular architecture, and effectively enhances the image information reconstruction quality in a low-illumination scene.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a low-light vision enhancement system and method for glaucoma patients. Background Technology

[0002] In recent years, digital image frequency domain enhancement technology, with its advantages of multi-scale decomposition and structural fidelity, has been widely used in computer vision and environmental perception systems under various complex lighting conditions. Especially in visual image enhancement scenarios for visually impaired individuals such as those with glaucoma, a typical image processing pipeline usually aims to solve a core computational problem: how to extract and enhance the structural details of environmental obstacles in real time from noisy video streams in extremely low light conditions (such as indoor nighttime, outdoor dusk, etc.) to compensate for the visual acquisition loss caused by the user's decreased contrast sensitivity and reduced dark adaptation ability.

[0003] However, in real-world applications with extremely low illumination, environmental images captured by cameras often exhibit extremely low signal-to-noise ratios, weak real edges, and are severely overwhelmed by high-frequency camera noise. Existing low-light frequency domain enhancement techniques often employ rigid decision criteria with static hard thresholds or relying solely on absolute signal energy within the frequency band when modulating image subband coefficients. Faced with extremely weak real structural signals, this rigid frequency domain filtering mechanism lacks the ability to perceive scene difficulty and cross-scale cross-validation: an excessively high decision threshold will cause weak edge signals to be judged as noise by the algorithm and directly erased, resulting in the loss of key spatial structures (such as steps, low obstacle outlines) in the reconstructed enhanced image; while relaxing the threshold cannot effectively isolate high-frequency background noise, resulting in the final output display image being filled with high-frequency artifacts, which seriously exacerbates visual interference.

[0004] Faced with extreme signal-to-noise ratio distributions in images under extremely low illumination, existing frequency domain enhancement methods still encounter many underlying bottlenecks in practical computation and applications. Therefore, there is an urgent need for an adaptive low-light image frequency domain enhancement system capable of multi-scale structural cross-validation and collaborative enhancement. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a low-light vision enhancement system and method for glaucoma patients.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In one aspect, this application discloses a low-light vision enhancement system for glaucoma patients, including a data acquisition module, a metric extraction module, a decision coupling module, and an image reconstruction module.

[0008] The data acquisition module acquires environmental images, bias parameters for adjusting brightness compensation, and weight coefficients for controlling signal classification;

[0009] The metric extraction module performs multi-scale frequency band decomposition on the environmental image to obtain multiple sub-band coefficients; it extracts the average amplitude of the sub-band coefficients and the noise amplitude corresponding to each frequency band, and calculates the sensing gap based on the ratio of the average amplitude to the noise amplitude and the bias parameter; it calculates the proportion of the square of the sub-band coefficient in the sum of the squares of the sub-band coefficients and the squares of the noise amplitude to obtain the signal saliency; it calculates the sign consistency decision results between the sub-band coefficients of adjacent frequency bands, and the ratio of the minimum amplitude to the maximum amplitude between the sub-band coefficients, and then obtains the cross-scale stability.

[0010] The decision coupling module uses the perception gap degree being greater than a preset gap degree threshold as a trigger gate, calculates the product of the difference between the perception gap degree and the gap degree threshold, the weight coefficient, the cross-scale stability, and the complement of the signal saliency, constructs a correction compensation term for each of the sub-band coefficients, and superimposes the correction compensation term onto the signal saliency to obtain the calibrated saliency.

[0011] The image reconstruction module constructs a mapping relationship between the perceived gap, the signal saliency, and the calibration saliency, then modulates each sub-band coefficient and obtains the enhanced environmental image through inverse transformation.

[0012] Secondly, this application discloses a low-light vision enhancement method for glaucoma patients, which is applied to the aforementioned low-light vision enhancement system for glaucoma patients, and includes the following steps:

[0013] Acquire environmental images, bias parameters for adjusting brightness compensation, and weight coefficients for controlling signal classification;

[0014] The environmental image is decomposed into multiple sub-band coefficients at multiple scales. The average amplitude of the sub-band coefficients and the noise amplitude corresponding to each frequency band are extracted. The perceptual gap is calculated based on the ratio of the average amplitude to the noise amplitude and the bias parameter. The signal saliency is obtained by calculating the proportion of the square of the sub-band coefficient in the sum of the squares of the sub-band coefficients and the squares of the noise amplitude. The sign consistency decision results between the sub-band coefficients of adjacent frequency bands and the ratio of the minimum amplitude to the maximum amplitude between the sub-band coefficients are calculated to obtain the cross-scale stability.

[0015] Using the sensing gap degree being greater than a preset gap degree threshold as a trigger gate, the product of the difference between the sensing gap degree and the gap degree threshold, the weight coefficient, the cross-scale stability, and the complement of the signal saliency is calculated to construct a correction compensation term for each of the sub-band coefficients, and the correction compensation term is superimposed on the signal saliency to obtain the calibrated saliency.

[0016] A mapping relationship is constructed between the perceived gap, the signal saliency, and the calibration saliency. Then, the coefficients of each sub-band are modulated and subjected to inverse transformation to obtain the enhanced environmental image.

[0017] Thirdly, this application discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the low-light visual assistance enhancement method for glaucoma patients as described in the second aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. This application breaks through the limitations of traditional rigid classification thresholds, achieving adaptive recovery of weak signals in challenging scenarios. Existing frequency domain enhancement techniques often use fixed signal saliency thresholds, leading to the easy misclassification and removal of real but weak edge signals as noise in extremely low-light scenarios. This application calculates the "perceived gap degree" through a metric extraction module, which is then used as a trigger gate by a decision coupling module. When the perceived gap degree exceeds a preset gap degree threshold (i.e., it is determined that there is recognition difficulty in the current local scene), the system can actively construct a correction compensation term and add it to the signal saliency, thereby obtaining a "calibrated saliency" after relaxing the standard. This mechanism introduces the "objective difficulty of the scene" into the underlying frequency domain signal classification logic, enabling the system to actively relax the classification standard in the dark areas where assistance is most needed, effectively extracting and preserving key structural details that are difficult to recover using traditional algorithms.

[0020] 2. This application introduces cross-scale features for cross-validation, balancing detail enhancement and high-frequency noise suppression. While relaxing the classification criteria for weak signals, this can easily lead to the side effect of amplifying high-frequency pure noise. Therefore, the sign consistency and amplitude ratio of sub-band coefficients between adjacent frequency bands are extracted to construct a "cross-scale stability." In the decision coupling module, cross-scale stability is directly involved in the calculation of the correction compensation term as a product term. This means that only when the weak signal has consistent physical structure support across multiple frequency bands (high stability) will it obtain significant compensation gain; isolated random noise (low stability) cannot be compensated. This design ensures accurate extraction of true edges in low signal-to-noise ratio environments without causing severe noise artifacts.

[0021] 3. This application constructs a sophisticated multidimensional self-consistent coupling mechanism, which greatly improves the robustness and reconstruction quality of image frequency domain modulation. A highly self-consistent product logic is employed: by introducing the "complement of signal saliency" into the product operation, it ensures that the correction compensation term only has a significant effect on pixels "originally judged as noise or extremely weak," while naturally attenuating the compensation strength for signals already possessing high saliency. This fundamentally avoids the repeated amplification or overexposure of the original clear signal. Combining the weighting coefficients and the perceptual gap difference, the system achieves non-trivial cross-coupling of environmental conditions, cross-scale structure, and the original signal state, providing an accurate frequency domain modulation reference for the final inverse image transformation reconstruction, significantly improving the information reconstruction quality of digital images under complex low-light conditions. Attached Figure Description

[0022] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0023] Figure 1 This is an overall framework diagram of the system described in this invention;

[0024] Figure 2 This is an overall framework diagram of the method described in this invention;

[0025] Figure 3 This is the overall flowchart of the present invention;

[0026] Figure 4 This is a flowchart illustrating the feedback update mechanism described in this invention. Detailed Implementation

[0027] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0028] In existing technologies, traditional low-light digital image frequency domain enhancement systems mostly employ static filtering algorithms with fixed thresholds or relying on absolute energy amplitude, making it difficult to balance the preservation of faint edge details with the suppression of high-frequency background noise in extremely low-light environments. While traditional signal saliency determination mechanisms can isolate obvious high-frequency noise, they lack the perception and adaptive relaxation mechanisms for the difficulty of local scenes. When the signal-to-noise ratio of the underlying image is extremely poor, they are prone to misjudging real but weak structural edges as noise and erasing them. Especially in complex environments with dim lighting and low contrast, existing systems typically process frequency band coefficients in isolation, lacking cross-validation of cross-scale structural features. This leads to severe aliasing and amplification of high-frequency sensor artifacts when attempting to forcibly enhance dark details. This processing method not only destroys the original spatial coherence of the image but may also cause visual confusion for users due to structural loss and artifact interference, greatly limiting the reliability of low-light image information reconstruction.

[0029] To address these issues, research has found the need to break away from traditional rigid absolute classification logic and establish a dynamic decision mechanism driven by environmental difficulty and possessing structural cross-validation. Particularly in extremely low signal-to-noise ratio (SNR) scenarios where real signals are severely obscured, the system needs to automatically perceive the degree of gaps in the local field of view, using this gap degree as a gating mechanism to adaptively relax the underlying frequency domain classification standard. Simultaneously, a cross-scale physical structural benchmark must be established to provide independent verification credentials for recovering weak signals, thereby achieving effective focusing on real edges and precise removal of pure noise. Based on this, this scheme derives a low-light vision-assisted enhancement system. By constructing a multi-dimensional feature network including perceived gap degree, signal saliency, and cross-scale stability, a decision coupling mechanism triggered by gap degree is established to generate calibration saliency. This achieves the goal of producing high-fidelity enhanced images from extremely low SNR image streams by dynamically calibrating the confidence of weak signals and accurately modulating frequency domain subband coefficients.

[0030] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] Example 1

[0032] like Figure 1 As shown in the figure, this embodiment introduces a low-light vision enhancement system for glaucoma patients, including the following modules: data acquisition module 100, metric extraction module 200, decision coupling module 300, and image reconstruction module 400.

[0033] The data acquisition module 100 is used to acquire environmental images, bias parameters for adjusting brightness compensation, and weight coefficients for controlling signal classification.

[0034] Specifically, this embodiment provides a low-light vision enhancement system for glaucoma patients, applied to a smart glasses platform. The smart glasses include at least a front-facing camera, an eye-tracking unit, an inertial measurement unit, an ambient light sensor, and a display screen. In this embodiment, the front-facing camera outputs grayscale video at 30fps, and the eye-tracking unit samples at a frequency of 60Hz. The system includes a data acquisition module, a metric extraction module, a decision coupling module, an image reconstruction module, a feedback update mechanism, and a memory for transferring state variables and inter-frame data.

[0035] The system operates using two asynchronous time-series links. The first is the video frame processing link, running at the video frame rate, responsible for sequentially performing metric extraction, decision coupling, and image reconstruction on each frame. The second is the feedback parameter update link, triggered by gaze events (occurring approximately 1 to 2 times per second), responsible for updating adaptive parameters based on user behavior. Data exchange between the two links is achieved through global state variables: the feedback link refreshes the state variable values ​​each time the update condition is met, and the video frame link reads the latest version of these variables during each frame processing. The two links do not require synchronization and there is no blocking relationship.

[0036] Specifically, the data acquisition module acquires three types of data: environmental images, bias parameters for adjusting brightness compensation, and weighting coefficients for controlling signal classification. The environmental image is a grayscale image of the current frame output by the front-facing camera of the smart glasses, denoted as... ,in and These are the horizontal and vertical pixel coordinates of the image, respectively. For video frame indexing, unless otherwise specified in this document, Both represent video frame indices. The offset parameter used to adjust brightness compensation is denoted as... , representing the individual offset, where the superscript For scale index, subscript The brightness range index is a global state variable asynchronously maintained by a feedback update mechanism. It is initialized to zero upon system startup and ranges from zero to 0.3. It accumulates only when the system detects a user's visual struggle signal, and slowly returns to zero through a decay mechanism when no struggle signal is detected. The weighting coefficient used to control signal classification is denoted as the weighting coefficient. Similarly, it is maintained asynchronously by the feedback update mechanism as a global state variable, and its initial value in this embodiment is denoted as . The value is set to 0.3, representing the baseline value of the weighting coefficient. At the start of each frame processing, the data acquisition module reads the current frame image output by the camera, as well as the current values ​​of the bias parameters and weighting coefficients after the most recent update by the feedback update mechanism.

[0037] The metric extraction module 200 is used to perform multi-scale frequency band decomposition on the environmental image to obtain multiple sub-band coefficients; extract the average amplitude of the sub-band coefficients and the noise amplitude corresponding to each frequency band; calculate the perceptual gap based on the ratio of the average amplitude to the noise amplitude and the bias parameter; calculate the proportion of the square of the sub-band coefficient in the sum of the square of the sub-band coefficient and the square of the noise amplitude to obtain the signal saliency; calculate the sign consistency decision result between the sub-band coefficients of adjacent frequency bands, and the ratio of the minimum amplitude to the maximum amplitude between the sub-band coefficients, and then obtain the cross-scale stability.

[0038] After data acquisition, the metric extraction module processes the environmental image, independently extracting initial metrics for subband coefficients from three orthogonal dimensions: perceptual asymmetry (environmental condition dimension), signal saliency (spatial location dimension), and cross-scale stability (cross-scale structure dimension). It should be understood that the actual execution order of intra-frame calculations in this module is as follows: first, frequency band decomposition is performed to obtain all original subband coefficients; then, the principal direction of each pixel is determined, and directional coefficients are obtained accordingly; subsequently, the scene components of the average amplitude and perceptual asymmetry of the edge region are calculated using the directional coefficients; then, the perceptual asymmetry is synthesized; and finally, signal saliency and cross-scale stability are calculated. The following description is organized according to logical dependencies to clearly demonstrate the physical meaning of each metric.

[0039] Specifically, a large-scale Gaussian low-pass filter is first applied to the grayscale image to obtain a local brightness field. A large-scale Gaussian low-pass filter refers to convolving the image with a two-dimensional Gaussian kernel whose standard deviation is greater than a preset smoothing threshold (2.0 in this embodiment) to extract the gradually varying brightness distribution of the image and suppress local texture details. Based on this brightness field, all pixels in the image are clustered into two brightness intervals according to their brightness values, and the interval indices are denoted as [insert indices here]. The range of values ​​is a set. These correspond to the dim and medium brightness areas, respectively, and the representative brightness value for each area is denoted as... . Frame The middle belongs to the brightness range The set of pixels is denoted as A soft transition weight in the form of a sigmoid is used between brightness ranges to avoid hard jumps at the range boundaries. In frame For brightness range The soft membership degree is denoted as The value is a real number between zero and one, and satisfies the following conditions: The soft membership degree is calculated using a sigmoid function centered on the boundary brightness value between the two brightness intervals.

[0040] Based on brightness field processing, further processing of grayscale images is performed. Perform directional multi-scale decomposition with downsampling. In this embodiment, a directional Laplacian pyramid structure is used, employing 3 directions and 4 scale layers. The directional indices for the 3 directions are denoted as... The range of values ​​is a set. These correspond to the 0°, 60°, and 120° directions, respectively. The scale indices of the four scale layers are denoted as... The range of values ​​is a set. ,in Corresponding to the highest frequency and highest resolution layer, This corresponds to the lowest frequency and coarsest scale layer. A total of 12 sub-band coefficient maps are generated, denoted as... , indicating direction ,scale Next pixel In frame The subband coefficient value. The spatial resolution of the sub-band is 1 / 3 of the original image resolution. times, that is Layers without downsampling maintain the original resolution. The layer is one-half. The layer is one-quarter. Each layer is one-eighth. The center frequency of each frequency band is denoted as... The design parameters of the decomposed filter bank are determined. This method incorporates a typical energy reference for the frequency band, denoted as . This constant is pre-configured before the system leaves the factory. Its value is calculated based on the mean absolute value of the coefficients of all images in the corresponding sub-band from a known natural image dataset (such as a public visual dataset containing various lighting conditions), reflecting the frequency band. The magnitude of typical coefficients in natural images is used to eliminate the inherent difference that the magnitude of high-frequency coefficients is naturally much smaller than that of low-frequency coefficients in subsequent cross-scale comparisons.

[0041] After completing the frequency band decomposition, it is necessary to determine the principal orientation of each pixel to ensure that subsequent cross-scale stability calculations are performed in the same physical direction. Specifically, the orientation-normalized total energy of the pixel is first calculated. For each pixel... Its direction The normalized total energy is denoted as The specific calculation formula is as follows:

[0042] ;

[0043] The above formula sums the absolute values ​​of the normalized coefficients for all four scale layers, using a typical energy benchmark of the frequency band for normalization. The purpose is to avoid the naturally large amplitude of low-frequency coefficients dominating the direction selection result, ensuring that the directional features of high-frequency details can also participate normally in the main direction determination. For scale layers... For cases where the value is greater than or equal to 2, the subband coefficients are determined using the nearest neighbor mapping method in full-resolution coordinates, i.e., low-resolution grid points are used. The coefficient value at the given location. Here, the nearest neighbor mapping is used instead of the neighborhood extremum mapping because the direction-normalized total energy is only used for comparing the relative strength between directions, the nearest neighbor accuracy is sufficient, and it avoids introducing cross-space neighborhood coupling before the main direction is determined.

[0044] Based on this, candidate main directions are determined, denoted as... The direction index, which maximizes the total energy after direction normalization, is expressed by the following formula:

[0045] ;

[0046] To suppress frequent jitter in the inter-frame direction, a hysteresis switching mechanism is introduced. Pixels in the frame... The final principal direction is denoted as In this embodiment, the hysteresis handover threshold ratio is denoted as . Taking 1.2 means the total energy in the new direction. Switching is only allowed when the energy exceeds 1.2 times that of the current direction. The determination rule is as follows:

[0047] when hour: ;

[0048] when hour: ;

[0049] in The main direction is determined by the previous frame. This is the normalized total energy calculated along the main direction of the previous frame in the current frame. This is relevant during a system cold start. The candidate main direction is directly taken as the main direction, that is .

[0050] After determining the main direction, define the pixel in the frequency band. The directional coefficient on is denoted as To retain the original sign (positive or negative), the specific formula is:

[0051] ;

[0052] in This represents the sub-band coefficient corresponding to the main direction. For coarser-scale layers ( (Greater than or equal to 2), the rule for determining the orientation coefficient in high-resolution coordinates is: first, map the pixel coordinates to low-resolution grid points. Then, within a 3×3 low-resolution neighborhood centered on the grid point, candidate coefficients with absolute values ​​less than a preset noise threshold are first eliminated. Then, the coefficient with the largest absolute value among the remaining candidate coefficients is selected and its original sign is retained. The preset noise threshold is 2.5 times the noise amplitude of the current frequency band. If no candidate coefficients exceeding the preset noise threshold exist within the 3×3 neighborhood, the center grid point coefficient is used; if the absolute value of the center grid point coefficient is still less than the preset noise threshold, the orientation coefficient corresponding to that scale layer is set to zero. This neighborhood corresponds to approximately [missing information - likely a missing section in the original image]. take In high-resolution regions, this neighborhood extremum approach absorbs the fractional pixel shift caused by downsampling, ensuring physical alignment during subsequent cross-scale comparisons. For scale layers... For all layers with a value greater than or equal to 2, the direction index used when determining the value is forced to be the main direction determined by the highest resolution layer for that pixel. This ensures that cross-scale comparisons are performed in the same physical direction.

[0053] Based on the determined principal direction and direction coefficients, the perceptual gap is further calculated. The perceptual gap quantifies the degree of visual information loss in various frequency bands and brightness ranges within the current scene. Its value ranges from zero to one; a value closer to one indicates that edge signals are more submerged by noise and the need for enhancement is more urgent, while a value of zero indicates sufficient signal and no enhancement is required. The perceptual gap employs an additive hierarchical structure of scene prior components and individual offsets: the scene components are calculated in real-time based on objective image quantities and can be used immediately upon system cold start; the individual offsets are obtained through long-term learning via behavioral feedback and are effective under all scene conditions.

[0054] The calculation of scene components requires first obtaining a frequency band noise reference. The standard deviation of the frequency band noise reference is denoted as... , indicating frequency band The noise amplitude introduced by the sensor is estimated by the automatic gain control and sensitivity status linkage of the camera module. It is a frequency band level noise quantity that is dynamically updated with the camera's operating status.

[0055] In one specific implementation, the frequency band noise reference is determined through a combination of dark frame calibration and runtime correction. During the dark frame calibration phase, multiple automatic gain states and sensitivity states are set under shading conditions, and at least 100 dark-field images are acquired. For each dark-field image, the same directional multi-scale decomposition as in this embodiment is performed, and the standard deviation of the sub-band coefficients at each scale level is calculated to obtain the initial noise amplitude at the corresponding automatic gain state, sensitivity state, and scale level. These initial noise amplitudes are stored in the system memory in the form of a lookup table.

[0056] During system operation, the data acquisition module reads the camera's current automatic gain and sensitivity values, and interpolates using the lookup table to obtain the base noise amplitude for the current frequency band. Simultaneously, it selects flat regions in the current frame where the gradient amplitude is below a preset flatness threshold and the local brightness variance is below a preset variance threshold. The robust standard deviation of the corresponding sub-band coefficients within these flat regions is then calculated and used as the runtime noise correction. The final frequency band noise amplitude is obtained by weighting the base noise amplitude and the runtime noise correction.

[0057] Specifically, the weight of the base noise amplitude can be 0.7, and the weight of the runtime noise correction can be 0.3. When the number of pixels in the flat region that meets the conditions in the current frame is less than 2% of the total number of pixels in the image, the confidence of the runtime noise correction is considered insufficient, and the base noise amplitude is directly used as the noise amplitude of the current frequency band.

[0058] When the camera's auto-gain or sensitivity values ​​are unavailable, the system employs an image self-estimation method to determine the frequency band noise amplitude: pixels with gradient amplitudes in the lowest 20% of the current frame are selected as candidate flat pixels. The median absolute deviation of the sub-band coefficients corresponding to each scale for these candidate flat pixels is calculated, and this median absolute deviation is converted into a noise amplitude in the form of standard deviation. This method provides a frequency band-level noise benchmark for signal saliency and perceptual asymmetry calculations even without camera parameter input.

[0059] It should be understood that, to avoid a circular dependency on the signal saliency of the current frame when calculating the perceptual gap, the signal saliency of the previous frame is used as the edge region mask. Optionally, before constructing the edge region mask of the current frame using the signal saliency of the previous frame, the signal saliency map of the previous frame is globally translated, rotated, or affine compensated based on the head angular velocity, attitude change, or global motion vector obtained based on image patch matching output by the inertial measurement unit, so that the edge mask of the previous frame is aligned with the image coordinates of the current frame. The complete calculation method of the signal saliency will be explained later. During the system cold start phase, i.e., within the first M frames, the entire image is not directly used as the edge region mask. Instead, candidate edge masks are generated based on the gradient magnitude or the magnitude of the highest frequency sub-band coefficient of the current frame. M is a preset number of frames, which is 5 in this embodiment. Specifically, pixels with gradient magnitude or the magnitude of the highest frequency sub-band coefficient in the current frame that are in the top 20% are selected as candidate edge pixels. After the number of current frames reaches M, the system switches to a method that determines the edge region based on the signal saliency of the previous frame and the sub-band response of the current frame.

[0060] Specifically, for the brightness range Internal frequency band The edge region is defined as the intersection of three sets of pixels: pixels whose signal saliency exceeds 0.5 in the previous frame, pixels belonging to the corresponding brightness range in the current frame, and pixels whose corresponding sub-band coefficient amplitude exceeds 2.5 times the noise amplitude of the current frame. This set is denoted as […]. The signal saliency is denoted as , The saliency of the signal in the previous frame is expressed by the following formula:

[0061] ;

[0062] Based on this, the average amplitude of the edge region is calculated and denoted as... In this embodiment, the clarity visibility coefficient is denoted as... The value of 3 indicates that a signal is considered sufficiently clear and discernible when its energy reaches three times the noise level; the minimum sample softening constant is denoted as... The value is set to 10 to eliminate the risk of division by zero in the empty set and to avoid jumps caused by fluctuations in sample size near the threshold. The formula for calculating the average amplitude of the edge region is:

[0063] ;

[0064] in This represents the total number of pixels in the set. When the sample size is sufficient, the softening constant has a negligible impact on the result, and the formula degenerates into the standard mean. When the sample size is small, the denominator is increased, and the average amplitude of the edge region is smoothed down. When the sample size is zero, the numerator is also zero, and the risk of division by zero is naturally eliminated. The entire process is continuous in response to changes in the sample size.

[0065] The scene component of perceived gap is denoted as superscript For scale index, subscript This is the brightness range index, with subscript 0 indicating that the quantity contains only the initial component of the scene prior. The calculation formula is:

[0066] ;

[0067] The formula combines the average amplitude of the edge region with the clear energy threshold. In comparison, and by performing a negative correlation inversion mapping, we obtain In this embodiment, the sharpness visibility coefficient is set to 3, meaning that when the average amplitude of the edge region reaches 3 times the amplitude of the band noise, the corresponding edge signal is considered to be discernible. When the average amplitude of the edge region is much smaller than the sharpness energy threshold, the scene component is close to one, indicating that the edge signal is completely submerged by noise; when the average amplitude of the edge region reaches or exceeds the sharpness energy threshold, the scene component is zero.

[0068] Furthermore, the perception gap is synthesized additively with scene components and bias parameters. The synthesized sensing gap is denoted as . The specific formula is as follows:

[0069] ;

[0070] The key advantage of using additive synthesis is that the bias parameter can still play a role when the scene component is zero (i.e., the objectively judged signal is clear in a bright scene), ensuring that the individual adaptability is effective under all scene conditions.

[0071] Then, the scalar perceptual notch at the brightness interval level is mixed into a pixel-wise spatial continuous field through pixel soft membership, denoted as... The specific formula is as follows:

[0072] ;

[0073] The resulting pixel-level perceptual gap field has a value between zero and one, and transitions smoothly at the boundary of the brightness range. This value is used as a unified input in the subsequent decision coupling module and image reconstruction module.

[0074] After calculating the sensing gap, the next step is to calculate the signal saliency. Signal saliency quantifies whether a real signal exists at the corresponding location of the sub-band coefficient. Its value ranges from zero to one; a value closer to one indicates a higher probability of a real image edge or texture structure at that location, while a value closer to zero indicates a higher probability of only sensor noise. The signal saliency is denoted as... The calculation formula using Wiener filtering is:

[0075] ;

[0076] in Let be the variance of the noise amplitude. This formula has a Wiener filtering form, approaching one when the signal energy is much greater than the variance of the noise amplitude, and approaching zero when it is much smaller than the variance of the noise amplitude. The signal saliency calculated in the current frame is used normally in subsequent steps of the same frame, and is also buffered for use as the edge mask input in the perceptual notch calculation in the next frame.

[0077] Furthermore, cross-scale stability is calculated. Cross-scale stability measures whether the signal at a pixel location has structural consistency across multiple scale levels, and its value ranges from zero to one. Real image edges produce responses at multiple scale levels with consistent signs and similar normalized amplitudes, thus exhibiting high cross-scale stability; pure noise, on the other hand, has independent responses at each scale level, resulting in low cross-scale stability.

[0078] To eliminate the cross-scale coefficient amplitude imbalance caused by the spectral regularity of natural images, typical energy normalization is first performed on the coefficients of each frequency band. The normalized directional coefficient is denoted as... The specific formula is as follows:

[0079] ;

[0080] After normalization, the typical amplitudes of each frequency band fall within the same order of magnitude. The initial cross-scale stability is denoted as... The final cross-scale stability is denoted as In this embodiment, the zero constant is denoted as Take a very small positive number, and the calculation formula is:

[0081] ;

[0082] in For frequency band The set of adjacent frequency bands, for internal frequency bands ( or ) contains two neighbors, for the boundary frequency band ( Only , Only It contains one neighbor; The number of elements in the set; Represents the normalized directional coefficients of adjacent frequency bands; This is an indicator function that takes the value 1 when the condition within the square brackets is true, and 0 otherwise. It should be understood that when any normalized coefficient is zero, the product of the two coefficients is not greater than zero, and the indicator function takes the value zero. This formula applies to frequency bands. For each of its neighboring frequency bands, the sign of the product of the two coefficients is first determined by whether they are consistent. Then, the minimum-to-maximum ratio of the normalized absolute values ​​is calculated to reflect the degree of amplitude similarity. The contribution of all neighbors is averaged.

[0083] To avoid random noise peaks from being included in cross-scale stability calculations, this embodiment sets a preset stability threshold for cross-scale stability. Only when the cross-scale stability exceeds the preset threshold does its corresponding stability contribution participate in the correction compensation term; when the cross-scale stability is lower than or equal to the preset threshold, the corresponding stability contribution is set to zero. In this embodiment, the preset stability threshold is 0.25. The structure compensation trigger condition is as follows: for the highest or second highest frequency sub-band, if the signal significance of the current frequency band is greater than 0.6, and the main direction of the current pixel remains consistent within two consecutive frames, it is considered that there is evidence of a single-scale high-frequency structure at that location. In this case, even if the response of its adjacent low-frequency sub-band is weak, the location is allowed to participate in subsequent correction compensation with a stability lower limit of 0.25 to preserve high-frequency details such as step edges and low-lying obstacle outlines.

[0084] The decision coupling module 300 is used to calculate the product of the difference between the perceived gap degree and the gap degree threshold, the weight coefficient, the cross-scale stability and the complement of the signal saliency, with the perceived gap degree being greater than a preset gap degree threshold as the trigger gate, to construct a correction compensation term for each of the sub-band coefficients, and to superimpose the correction compensation term onto the signal saliency to obtain the calibrated saliency.

[0085] After extracting the initial metrics of the above three dimensions, the specific implementation of the decision coupling module will be described below.

[0086] The decision coupling module uses a perceived gap value greater than a preset gap value threshold as a trigger gate to construct correction compensation terms for each sub-band coefficient. These correction compensation terms are then superimposed on the signal saliency to obtain the calibrated saliency. This mechanism introduces scene difficulty into the frequency domain signal classification process—actively and conditionally relaxing the classification criteria between signals and noise in difficult scenes, thus achieving dynamic adaptive adjustment of the signal classification criteria based on the current scene difficulty characteristics.

[0087] Specifically, in this embodiment, the notch threshold is denoted as... The value is set to 0.3. The calibration significance is denoted as... The calculation formula is:

[0088] ;

[0089] This formula adds a correction compensation term to the signal saliency. The correction compensation term consists of the product of a weighting coefficient and three modulation factors, which respectively represent three gating conditions.

[0090] First modulation factor This is the difference between the perceived gap degree and the gap degree threshold. When the perceived gap degree exceeds the gap degree threshold, this factor is positive, and the correction compensation term adjusts the signal saliency upward, that is, relaxes the classification criteria, so that more edge signals are preserved in difficult scenarios.

[0091] Second modulation factor The complement of the signal significance, inverse mapping refers to the mapping of the signal significance... Mapped to complement The purpose of this operation is to correct and suppress strong signals, ensuring that the correction effect is concentrated only on locations with low signal saliency, i.e., those pixels that were originally judged to be close to noise. The first attenuation result... This represents the difference in notch strength after reverse suppression of the original signal strength. It ensures that the compensation effect of the notch strength is maximized only when the signal saliency is weak, avoiding overcompensation for already strong signals. For locations with high signal saliency, this factor is close to zero, with minimal correction, avoiding repeated amplification of already confirmed signals. In simple scenarios, since the perceived notch strength does not exceed the notch strength threshold, the correction compensation term is zero, and the system maintains the original signal saliency determination result, avoiding additional suppression of weak but identifiable edge signals.

[0092] The third modulation factor is cross-scale stability, providing a second opinion independent of the perceived gap and signal saliency. It ensures that only those locations with weak signals but supporting evidence of multi-scale structure can receive saliency boosts, acting as a safety valve to prevent purely noisy locations from being incorrectly boosted due to scene difficulties.

[0093] Three conditions must be met simultaneously for a significant positive correction to occur: the scenario must be genuinely difficult, the location must have originally been classified as a weak signal, and there must be evidence of cross-scale structure. This three-condition gating design makes the adjustment of signal classification criteria both proactive and conditional; when the perceived gap is less than or equal to the gap threshold, the correction compensation term is zero, and the original signal significance is not negatively suppressed.

[0094] The image reconstruction module 400 is used to construct the mapping relationship between the perceived gap, the signal saliency and the calibration saliency, and then modulate each sub-band coefficient and obtain the enhanced environmental image through inverse transformation.

[0095] Based on the calibration saliency output by the decision coupling module, the image reconstruction module further constructs a mapping relationship between the perceptual gap, signal saliency, and calibration saliency, and then modulates the coefficients of each sub-band, and obtains the enhanced environmental image through inverse transformation. Specifically, this mapping relationship is established layer by layer through the following steps.

[0096] First, the ideal compensation gain is calculated. To avoid indiscriminate amplification of noise in purely dark or structureless scenes, this embodiment sets up a low-information scene protection mechanism. Specifically, the proportion of candidate edge pixels and the global sub-band energy of the current frame are calculated. When the proportion of candidate edge pixels is less than 1% and the global sub-band energy is less than 1.5 times the frequency band noise variance, the current frame is determined to be a structureless scene. In structureless scenes, the upper limit of the ideal compensation gain is limited to 1.3 to suppress pure noise amplification. In other scenes, to eliminate the risk of divergence when the perceptual asymmetry approaches one when directly taking the reciprocal form, a hyperbolic tangent saturation mapping is used to convert the perceptual asymmetry into the ideal compensation gain, denoted as . In this embodiment, the maximum additional gain coefficient is denoted as... Take 6; the kurtosis coefficient of the saturation curve is denoted as . Let's take 2, and the calculation formula is:

[0097] ;

[0098] Where tanh is the hyperbolic tangent function. When the sensing gap approaches one, the ideal compensation gain approaches... The value is bounded; when the perceived gap approaches zero, the ideal compensation gain approaches 1, and no enhancement is performed.

[0099] Then, the enhancement demand degree is calculated. The enhancement demand degree is generated by the cross-generation of the perception gap degree and signal saliency, representing the necessity of enhancement at a certain location in a certain frequency band. The enhancement demand degree is denoted as... In this embodiment, the demand bias coefficient is denoted as... We take 0.65. The calculation formula is:

[0100] ;

[0101] Using asymmetric exponentiation prioritizes the weights of necessity over signal existence—the exponentiation. Greater than Even in challenging scenarios with weak signals, a trial enhancement is offered, with risks controlled through subsequent cost assessments. The value ranges from zero to one.

[0102] Next, the enhancement cost is calculated to assess the risk that the enhancement operation may introduce noise amplification. The enhancement cost is denoted as... The calculation formula is:

[0103] ;

[0104] First item To represent the calibrated noise percentage, the complement of the calibration significance is used here instead of the complement of the signal significance. This ensures that the cost assessment and the demand assessment use the same set of signal classification results calibrated for scenario difficulty, maintaining logical consistency. (Second item) As a pressure factor, its value ranges from 0.5 to 1, forming an endogenous negative feedback mechanism between demand and cost, preventing the system from being over-enhanced under high-demand scenarios.

[0105] Based on this, the initial gain weights are calculated, converging the enhancement demand, calibration significance, and enhancement cost into a unified gain decision. The initial gain weights are denoted as... In this embodiment, the numerical stability constant is denoted as . Take 0.001. The calculation formula is:

[0106] ;

[0107] The formula has a competitive fractional form: the gain numerator is the product of the enhancement demand and the calibration significance, representing the gain of the enhancement; the gain denominator represents the gain plus the enhancement cost. When the calibration significance approaches zero, the numerator approaches zero, and the initial gain weight automatically suppresses the gain; when the calibration significance is at a moderate level, the influence of the enhancement demand is attenuated with a linear weight, suppressing pure noise without excessively damaging weak signals. Spatial selectivity naturally emerges from this competitive structure, eliminating the need for a separate structure detection module.

[0108] Furthermore, cross-band conditional propagation correction is performed, using coarse-scale (low-frequency, high-confidence) judgment results to correct the gain decision at the fine-scale (high-frequency, susceptible to noise contamination). The propagation direction shifts from the coarse-scale to the fine-scale, i.e., indexed by scale. Layer Layered transfer is the standard paradigm for multi-scale signal processing—coarse-scale coefficients come from stronger low-pass filtering and greater spatial support, resulting in a more complete and reliable noise averaging effect.

[0109] The propagation-corrected gain weight is denoted as The propagation corrections are processed sequentially from coarse to fine. The coarse-scale layer data referenced in each rule are the initial gain weights, calibration significance, and signal significance before propagation correction, rather than the values ​​after propagation correction, to ensure that propagation only occurs in single-hop transmission between adjacent scales and does not produce cascading cumulative effects.

[0110] For the coarsest scale layer There is no coarser reference layer, so it directly makes For the remaining scale layers, in this embodiment, the first preset threshold is denoted as... The value is set to 0.15. Propagation judgment is only entered when the absolute value of the difference between the initial gain weights of adjacent frequency bands exceeds the first preset threshold; otherwise, the gain weight after propagation correction is directly set to equal the initial gain weight before propagation correction.

[0111] After entering the propagation judgment stage, the correction method is determined based on two mutually exclusive rules. In this embodiment, the propagation coefficient is denoted as... Take 0.3; the second preset threshold is , take 0.3; low frequency support threshold is denoted as Take 0.5; False support criterion is recorded as The value is set to 0.3; the inhibition coefficient is denoted as... Take 0.4.

[0112] Rule 1 is a coarse-scale confirmation propagation, triggered if and only if the initial gain weight of the low-frequency band is greater than the second preset threshold and the signal salience of the low-frequency band is greater than the low-frequency support threshold. The propagation correction formula is:

[0113] ;

[0114] in, This represents the positive gain correction amount. This represents the calibration significance of the adjacent coarse-scale layer j+1. This represents the initial gain weights of the adjacent coarse-scale layer j+1. To obtain a positive value, the larger of the value within the parentheses and zero is taken. Rule 1 uses the calibrated significance as the modulation factor for propagation intensity because Rule 1 requires a coarse-scale criterion for certain signal existence, and the calibrated significance better reflects the credibility of the true signal after correction for scene difficulty.

[0115] Rule 2 is spurious fine-scale suppression, triggered only if the initial gain weight of the high-frequency band is greater than the second preset threshold, the signal significance of the low-frequency band is less than the spurious support criterion, and the cross-scale stability of the high-frequency band is less than the preset high-frequency structure threshold. If the high-frequency band has evidence of a single-scale high-frequency structure or the cross-scale stability is not lower than the preset high-frequency structure threshold, then spurious fine-scale suppression is not performed. In this embodiment, the preset high-frequency structure threshold is 0.25. The suppression correction formula is:

[0116] ,in It is a negative inhibitory factor. Indicates the signal significance of coarse-scale layer j+1;

[0117] Rule 2 uses signal significance (rather than calibration significance) as the criterion to prevent weak signals recovered by the correction compensation term from being mistakenly regarded as strong supporting evidence.

[0118] Rule 1 requires the coarse-scale signal significance to be greater than the low-frequency support threshold of 0.5, while Rule 2 requires the coarse-scale signal significance to be less than the false support criterion of 0.3. The triggering conditions of the two rules are strictly mutually exclusive. When neither rule is met, no correction is made, and the gain weight after propagation correction is equal to the initial gain weight before propagation correction.

[0119] The final gain ratio is denoted as The calculation formula is:

[0120] ;

[0121] The formula uses 1 (no enhancement) as the basis and interpolates between 1 and the ideal compensation gain according to the gain weight after propagation correction.

[0122] After obtaining the final gain ratio, gain application and image reconstruction are performed. The final gain ratio is calculated in a high-resolution coordinate system, while the native resolution of each subband coefficient varies with the scale layer. Therefore, before applying the gain to the subband coefficients, the gain map needs to be downsampled to the native resolution of the corresponding subband. Specifically, this is done by downsampling the high-resolution gain map to each subband's native resolution. Take the arithmetic mean within the pixel block to obtain the first... The gain map at the native resolution of the layer is denoted as ,in For the first The native low-resolution coordinates of the sub-band. For the highest resolution layer ( (without downsampling), the gain map after downsampling is the same as the high-resolution gain map.

[0123] Gain application performs pixel-wise multiplication on the subband coefficients for each direction and each scale layer; the enhanced subband coefficients are denoted as... The specific formula is as follows:

[0124] ;

[0125] It should be understood that here The first output of the frequency band decomposition The values ​​of the subband coefficients in their native low-resolution coordinates are used to indicate that the decomposition operation is applied to the full-resolution input image, and the output resolution of each subband decreases with the scale layer.

[0126] Regarding the gain application strategy, the final gain multiplier is obtained based on the analysis of the pixel's principal direction coefficients, but it is applied equally to all three sub-bands. The rationale for this design is that: on real edge pixels, the principal direction response is much larger than the non-principal direction response, and the relative error when the non-principal direction coefficients are amplified is limited; on noise-dominant pixels, the initial gain weight is close to zero, the final gain is close to 1, and the coefficients of each direction remain essentially unchanged. Therefore, the equal application design does not produce perceptible artifacts in typical scenarios and avoids the directional selectivity discontinuity problem that may be caused by direction-dependent gains.

[0127] Perform the inverse transformation of the decomposition (i.e., the inverse transformation of the directional Laplacian pyramid) on all the enhanced subband coefficients to reconstruct the enhanced image. The signal saliency of the current frame is buffered and used as the edge mask input in the perceptual gap calculation by the metric extraction module in the next frame.

[0128] The above describes the complete execution process of each step in the video frame processing chain within each frame. Asynchronously, the feedback update mechanism dynamically updates the bias parameters and weight coefficients in the state storage unit based on user behavior signals for the data acquisition module to read. The specific implementation of the feedback update mechanism is described below.

[0129] The feedback update mechanism first performs gaze behavior acquisition and gating. The smart glasses' built-in eye-tracking unit outputs a gaze point sequence at a frequency of 60Hz, the inertial measurement unit outputs walking speed, head angular velocity, and posture data, and the ambient light sensor outputs the current ambient illuminance, denoted as . Optionally, to meet the real-time processing requirements of the wearable smart glasses platform, the multi-scale decomposition, saliency calculation, cross-scale stability calculation, and gain application are all performed in a frame-by-frame pipeline manner; for coarser scale layers, intermediate results are calculated and cached on their native low-resolution grid; for the gain map, scale matching is performed only once before reconstruction. Optionally, the system performs full enhancement only on the image region within a preset viewing angle around the gaze point, and directly reuses the gain map from the previous frame for the peripheral region, thereby reducing the computational load.

[0130] The gaze point sequence is segmented using a velocity threshold recognition algorithm. In this embodiment, the velocity threshold is set to 30 degrees per second to distinguish between fixation and saccades, resulting in a gaze event sequence. Each gaze event is denoted as... subscript For the event sequence number, and The spatial mean of the coordinates of the fixation point during fixation. The duration of gaze. This is the moment the event begins.

[0131] Considering that elderly patients with low vision often exhibit typical obstacle avoidance behaviors in dimly lit and unfamiliar environments, such as stopping, turning their heads in place, and prolonged probing gaze, gating solely with a walking speed threshold would miss the most crucial stationary search scenarios. Therefore, this method employs extended visual search state gating, using parallel conditions of either a moving obstacle avoidance state or a head saccade state to filter effective gaze events. Condition A is the moving obstacle avoidance state: walking speed greater than 0.3 meters per second. Condition B is the head saccade state: walking speed less than or equal to 0.3 meters per second, but with a significant change in head angular velocity within the past 2 seconds. In this embodiment, the specific criterion is that the time period with a heading angle change exceeding 15 degrees per second accounts for more than 30%. If either condition is met, the gaze event proceeds to subsequent analysis. Condition B captures the stop-and-go visual search behavior while simultaneously excluding stationary single-point fixation by requiring active head saccades.

[0132] After gating and filtering gaze events, the selected events need to be attributed to their corresponding dominant frequency band and dominant brightness range. To eliminate attribution bias caused by micro-eye movements, head swaying, or scene movement in a single frame, this method performs temporal mean pooling within the gaze time window.

[0133] Specifically, an image window centered on the gaze point coordinates and corresponding to a 1-degree field of view is defined as... The dominant frequency band is denoted as The specific formula for the frequency band index that maximizes the normalized response after time-mean pooling is as follows:

[0134] ;

[0135] in This is the number of video frames during the gaze period, approximately equal to 30 multiplied by the dwell time. For For low-resolution layers with a value greater than or equal to 2, the orientation coefficients are taken in the same way as the rules defined in the metric extraction module, i.e., the full-resolution representation is obtained by using neighborhood extremum mapping.

[0136] The dominant brightness range is denoted as To provide an index for brightness ranges that cover more pixels during fixation, the specific formula is as follows:

[0137] ;

[0138] in For video frames The area within the central gaze window belongs to the brightness range. The number of pixels. The duration of the event is recorded in a sliding window of a two-dimensional grid defined by the dominant frequency band and the dominant brightness range.

[0139] Based on event attribution, a baseline for expected fixation duration needs to be established to distinguish between normal fixation and abnormally long fixation. The baseline duration for expected fixation duration is denoted as... This reflects the user's frequency band Brightness range Natural fixation duration under certain conditions.

[0140] During system cold start, the baseline is initialized using population prior values, denoted as . The group prior value is obtained by statistically analyzing the typical fixation durations of healthy elderly individuals and glaucoma patients under various frequency bands and brightness combinations, and is built into the system protocol. After system startup, an initial individual calibration period begins, consisting of 10 minutes or 300 cumulative valid fixation events, whichever comes first. During this initial individual calibration period, the system collects stable saccades from users at a lower learning rate and uses the median or truncated mean to individually correct the group prior; simultaneously, the bias parameter update learning rate is reduced to 20% of the normal learning rate.

[0141] In one specific implementation, the group prior values ​​are used to construct a two-dimensional baseline table based on frequency bands and brightness ranges. For four scale layers and two brightness ranges, the two-dimensional baseline table includes eight baseline durations. The baseline duration is in seconds and is used to represent the duration of a user's natural gaze within the corresponding frequency band and brightness range.

[0142] In this embodiment, the group prior baseline duration can be set as follows: 0.85 seconds for the highest frequency scale layer and 0.65 seconds for the dim and moderate brightness regions; 0.95 seconds for the second highest frequency scale layer and 0.70 seconds for the dim and moderate brightness regions; 1.10 seconds for the mid-to-low frequency scale layer and 0.80 seconds for the dim and moderate brightness regions; and 1.25 seconds for the lowest frequency scale layer and 0.90 seconds for the dim and moderate brightness regions. It should be understood that the above baseline durations are initial system values ​​and do not limit the scope of protection of this invention; the baseline durations can be adaptively corrected through an initial individual calibration period under different ages, visual abilities, walking speeds, or usage scenarios.

[0143] During the initial individual calibration period, the system does not directly update the bias parameters based on a single long fixation event. Instead, it statistically analyzes the duration of fixation events during steady saccades and calculates the truncated mean or median for each frequency band and brightness range. If the number of valid fixation events collected in a certain frequency band and brightness range during the initial individual calibration period is no less than 20, the individual statistical value corresponding to that frequency band and brightness range is used to replace the group prior baseline duration. If the number of valid fixation events is less than 20, the group prior baseline duration is retained and continues to be slowly updated using an exponential moving average method during subsequent operation.

[0144] To avoid contaminating the baseline due to non-visually challenging events such as abnormal pauses, conversations, or looking up at distant objects, fixation events included in baseline statistics must simultaneously meet the following conditions: walking speed greater than 0.3 m / s or head angular velocity within the normal saccade range during the fixation event; no sudden stops, significant turns, or abrupt drops in stride length within 500 ms after the fixation event ends; and the final gain factor of the current frame is less than 1.5. Fixation events meeting these conditions are considered stationary saccade events and are used to update the expected fixation duration baseline.

[0145] The baseline is continuously updated dynamically during operation. When the user is in a safe sampling state, the collected steady-state saccade samples (state C events defined in the cognitive state scoring section below) are used to update the baseline. The safe sampling state preferably meets two conditions simultaneously: first, the inertial measurement unit indicates that the user is in a normal walking state, i.e., the walking speed is greater than 0.3 meters per second; second, the ambient illuminance exceeds a preset sufficient threshold, which is 300 lux in this embodiment. If the system fails to collect steady-state saccade samples that meet the 300 lux condition for 72 consecutive hours, steady-state saccade samples are allowed to be collected under the conditions that the ambient illuminance exceeds 100 lux, the user is in a low-risk static scene, and the final gain ratio of the system is less than 1.5, to slowly update the expected fixation duration baseline; at this time, the baseline learning rate is 50% of the normal learning rate. The first condition ensures that the dynamic behavior pattern of the baseline is consistent with the actual runtime, and the second condition ensures that there is no visual enhancement requirement in the collection scene.

[0146] The baseline update uses an exponential moving average method. In this embodiment, the baseline learning rate is denoted as . The value is 0.002. When a new steady-state scan event is acquired, its dwell time is recorded as . At that time, the baseline is updated according to the following formula:

[0147] ;

[0148] in This indicates the update assignment of state variables.

[0149] Optionally, the baseline includes a timeliness protection mechanism. In this embodiment, the maximum allowable interval without updating is denoted as... The learning period is 72 hours. If the baseline is more than this interval from the last update, an update is performed with a larger learning rate. In this embodiment, the larger learning rate is denoted as _____. If we take 0.01, even if the illumination conditions for safe sampling are not met, we are still required to be in a normal walking state to prevent the baseline from becoming completely invalid due to a long-term lack of sufficient illumination sampling conditions.

[0150] It should be understood that visual struggle events and successful confirmation events, as defined in the cognitive state scoring section below, do not participate in baseline updates: the former prolongs fixation due to scene difficulty, and the latter prolongs fixation after clear vision and confirmation; neither reflects the natural baseline without enhancement needs. Only smooth saccadic events participate in updates. Temporally, cognitive state determination uses the latest baseline value, and the next round of baseline updates occurs after the current event's state determination is completed, forming a unidirectional temporal sequence without circular dependencies.

[0151] Once the baseline is established, a composite cognitive state score can be performed on each fixation event. This score is used to avoid statistically confounding long fixations caused by two different reasons: prolonged fixation due to unclear vision and confirmatory fixation after clear vision.

[0152] Specifically, the gait change detection method is as follows: within a preset duration interval after the gaze event ends (500 milliseconds in this embodiment), the statistical variance or the relative difference between the mean values ​​of the three inertial motion features output by the inertial measurement unit (IMU)—step frequency, step length, and heading angle—is calculated as the rate of change. When any rate of change exceeds the corresponding preset threshold, it is determined that there has been a significant adjustment in gait. In this embodiment, the threshold for the rate of change of step frequency is 20%, the threshold for the rate of change of step length is 25%, and the threshold for the rate of change of heading angle is 15 degrees per second. A gait struggle indicator variable is defined. In observing events The value is 1 when the gait stops or slows down in the posterior window after the end, and 0 when the gait is decisively adjusted.

[0153] The threshold for determining long fixation is denoted as In this embodiment, it is defined as 1.2 times the current baseline duration: ;

[0154] Cognitive weight is denoted as Based on the combination of fixation duration and gait pattern, fixation events are divided into three states:

[0155] when and hour: ;

[0156] when and hour: ;

[0157] when hour: ;

[0158] in and These represent the dominant frequency band and dominant brightness range for attributing the event. The first scenario corresponds to state A, the visual struggle state, where the fixation duration exceeds the long fixation threshold and the gait shows stagnation or deceleration, indicating that the user is prolonging their fixation and hesitating due to unclear vision. The second scenario corresponds to state B, the successful confirmation state, where the fixation duration exceeds the threshold but the gait is decisively adjusted, indicating that the user has clearly seen the target. The third scenario corresponds to state C, the smooth saccade state, covering all other scenarios.

[0159] The weight of state B is set to zero instead of negative. The rationale behind this design is that state B physically signifies sufficient rather than excessive enhancement. Giving it a negative weight would artificially drive a decrease in the bias parameter, potentially triggering oscillations between upward and downward adjustments. A weight of zero means that state B does not generate feedback driving force. When there is no struggle signal, the system slowly returns to a neutral configuration through the natural decay of the leak integrator, resulting in a clean design where struggle drives upward adjustments and decay drives downward adjustments. Although state B events do not participate in feedback parameter updates, their occurrence frequency is included in the overall system success rate statistics and can be used as a utility evaluation indicator output to users or medical personnel.

[0160] After obtaining the cognitive weights, the feedback bias is calculated. In this embodiment, the duration of the feedback bias statistical window is denoted as... Take 5 minutes. Regarding the frequency band... and brightness range For a given cell, the set of sequence numbers of all gaze events attributed to that cell within the time frame of the window preceding the current moment is denoted as . The cumulative number of events within the window is recorded as follows: This includes events for all states. When there are no events in the window, the feedback bias of that cell is zero, and the leaked integral decay update is still performed, so that the bias parameter gradually returns to zero when there is no feedback drive.

[0161] Feedback deviation is denoted as The calculation formula is:

[0162] ;

[0163] Wherein, the zero constant for prevention is denoted as (In this embodiment, the value is 0.001). Since the cognitive weights of states B and C are both zero, only visual struggle events contribute to the summation term. However, the denominator includes all events, so the feedback bias reflects both the frequency of struggle events and the severity of a single struggle. When the proportion of struggle events in a cell is low, the feedback bias is diluted and becomes smaller, avoiding over-adjustment caused by rare struggle events. The feedback bias is always non-negative.

[0164] Preferably, to reduce misjudgments caused by a single abnormal gaze event, this embodiment does not directly drive the update of feedback parameters based on a single visual struggle event. For a grid with a defined frequency band and brightness range, the feedback deviation only participates in the update of the bias parameters and weighting coefficients when the number of visual struggle events attributed to that grid within a 5-minute sliding window is no less than 3, and the average perceptual asymmetry of the image region corresponding to that grid exceeds 0.4.

[0165] After obtaining the feedback deviation of each cell, the bias parameters and weight coefficients are updated respectively.

[0166] The bias parameters are updated using a leak integrator. In this embodiment, the individual bias learning rate is denoted as... The value is set to 0.005; the forgetting decay coefficient is denoted as... Take 0.001. Update the formula as follows:

[0167] ;

[0168] in Represented as an incremental driving component, The formula represents the attenuation reference component. It adds an attenuation term to the conventional pure integral update, forming a leak integrator to prevent integral saturation. The lower cutoff is zero, and the upper cutoff is 0.3. The lower cutoff of zero is consistent with the mathematical properties of unidirectional feedback: since the feedback deviation is always non-negative, the bias parameter always remains non-negative and bounded, and gradually decays through the leak integrator term when there is no feedback drive. The lower cutoff of zero is precisely the target value for attenuation regression when there is no struggle signal for a long period. To avoid a sudden increase in parameters due to a single misjudgment, the maximum increment of the bias parameter within each sliding window is set to 0.01. When no visual struggle event is detected for three consecutive sliding windows, the forgetting attenuation coefficient is increased from 0.001 to 0.005 to accelerate the return of the bias parameter to a neutral configuration.

[0169] The dynamic logic of the update mechanism is as follows: When a struggle signal appears, the feedback deviation is greater than zero, the bias parameter is slowly increased, resulting in an increase in the perception gap, and the enhancement is more aggressive when the same conditions are encountered again; when there is no struggle signal, the feedback deviation is close to zero, the bias parameter is slowly decreased through leakage integral decay, and the system automatically returns to a conservative configuration to avoid long-term over-enhancement.

[0170] The weight coefficients are also updated using a leak integrator, but this is done at the global level. In this embodiment, the learning rate for the weight coefficients is denoted as... The value is set to 0.02; the forgetting decay coefficient of the weighting coefficient is denoted as... The value is set to 0.001; the preset activity threshold is denoted as... Take 5.

[0171] The set of active cells is denoted as , defined as the set of all cells within the current sliding window whose cumulative event count exceeds a preset activity threshold, and the number of elements in the set is denoted as . When the set of active cells is empty (i.e., there are no active cells in the current window), the global feedback feature is set to zero, and the steady-state regression update of the weight coefficients is still performed, so that the weight coefficients gradually return to their initial values.

[0172] When the active cell set is not empty, the global feedback feature is denoted as... , which is the arithmetic mean of the feedback biases of all active cells:

[0173] ;

[0174] The formula for updating the weight coefficients is:

[0175] ;

[0176] in For the steady-state regressive force term to the neutral value, This represents the deviation component. Expanding the right-hand side of the update formula yields... It can be seen that in the absence of struggle signals, the update degenerates into... This describes the process of the decay factor regressing to its initial value through the incomplete integration process. The lower cutoff is 0.3 (initial value) and the upper cutoff is 0.6. Since the global feedback feature is always non-negative, the weight coefficient cannot mathematically fall below its initial value. Aligning the lower cutoff with the initial value eliminates the mathematical dead zone. If there is no struggle signal in a certain period, the weight coefficient slowly regresses to its initial value through incomplete integration to reach a steady state, corresponding to the basic weak signal recovery strength. To avoid the global weight coefficient rising rapidly due to short-term misjudgment, the maximum increment of the weight coefficient within each sliding window is set to 0.02. When no visual struggle event is detected in three consecutive sliding windows, the forgetting decay coefficient of the weight coefficient increases from 0.001 to 0.005.

[0177] The bias parameter and weighting coefficient are naturally separated at the data consumption level: the bias parameter consumes the feedback deviation on a grid-by-grid basis and makes local adjustments to the corresponding frequency band and brightness range; the weighting coefficient consumes the global feedback feature quantity aggregated across grids and adjusts the global strength of the correction compensation term in the calibration significance. The two can achieve causal separation without active perturbation.

[0178] Optionally, to prevent sudden global attention shifts from causing abnormal updates to feedback parameters, a global consistency protection mechanism is set. In this embodiment, the threshold for drastic changes is denoted as... A value of 0.5 indicates that the average excess dwell time of struggle events exceeds 50% of the baseline; the light variation threshold is denoted as... The threshold is set at 50%. If, within a certain time period, the feedback deviation of more than 70% of the active cells simultaneously exceeds the threshold for drastic changes, the illuminance change rate of the ambient light sensor, the current average perceived asymmetry, and the current image signal-to-noise ratio are further checked. If the illuminance change rate does not exceed 50%, the current average perceived asymmetry is less than 0.4, and the current image signal-to-noise ratio is not less than 3dB, it is determined to be caused by global attention shift, and all parameter updates in this round are frozen. If the illuminance change rate exceeds 50%, or the current average perceived asymmetry is not less than 0.4, or the current image signal-to-noise ratio is less than 3dB, it is determined to be a real visual difficulty scene, and the freeze is waived, and parameter updates are performed normally.

[0179] The bias parameter, weighting coefficients, and expected gaze duration baseline are three parameter groups maintained asynchronously as global state variables. The video frame processing chain reads these variables during the processing of each frame: the metric extraction module reads the bias parameter for additive synthesis when calculating the perceptual aberration, and the decision coupling module reads the weighting coefficients to construct the correction compensation term. The feedback update mechanism reads the expected gaze duration baseline for calculating the long gaze threshold during each gaze event. The signal saliency of the current frame is buffered between frames and used by the metric extraction module in the next frame as the edge mask input in the perceptual aberration calculation.

[0180] The aforementioned closed loop forms a four-speed adaptive architecture: the first speed is the immediate response of scene priors, which adjusts the scene component of the perceived gap in real time based on the objective brightness and noise characteristics of each frame of image on a millisecond timescale; the second speed is the weight coefficient tracking of global struggle, which adjusts the global strength of the correction compensation term on a minute timescale; the third speed is the slow learning of bias parameters, which accumulates the visual ability deviation of individual users on an hour to day timescale; and the fourth speed is the lifelong adaptation of the expected gaze duration baseline, which tracks the long-term changing trend of user visual function on a week to month timescale.

[0181] It should be noted that the clear visibility coefficient, notch threshold, stability threshold, weight coefficient benchmark value, learning rate, typical energy benchmark of the frequency band, and population prior baseline duration given in this embodiment are all preferred implementation parameters. These parameters can be obtained from the system's factory calibration or through online statistics during the initial operation phase. As long as each parameter is used to implement the sensing notch calculation, signal saliency calculation, cross-scale stability calculation, calibration saliency generation, and feedback parameter update described in this embodiment, it constitutes an equivalent implementation of the technical solution of this invention.

[0182] The settings of the above parameters do not require reliance on a specific database or specific patient samples; when there is no pre-collected population sample, the system can start with default parameters and gradually update the bias parameters, weight coefficients and baseline duration based on steady saccade events and visual search feedback events during actual user use, thereby achieving individualized adaptation.

[0183] Example 2

[0184] like Figure 2 As shown, this embodiment introduces a low-light vision enhancement method for glaucoma patients, which is applied to the low-light vision enhancement system for glaucoma patients in Embodiment 1, and includes the following steps:

[0185] S1. Acquire environmental images, bias parameters for adjusting brightness compensation, and weight coefficients for controlling signal classification;

[0186] S2. Perform multi-scale frequency band decomposition on the environmental image to obtain multiple sub-band coefficients; extract the average amplitude of the sub-band coefficients and the noise amplitude corresponding to each frequency band, and calculate the sensing gap based on the ratio of the average amplitude to the noise amplitude and the bias parameter; calculate the proportion of the square of the sub-band coefficient in the sum of the square of the sub-band coefficient and the square of the noise amplitude to obtain the signal saliency; calculate the sign consistency decision result between the sub-band coefficients of adjacent frequency bands, and the ratio of the minimum amplitude to the maximum amplitude between the sub-band coefficients, and then obtain the cross-scale stability.

[0187] S3. Using the sensing gap degree being greater than a preset gap degree threshold as a trigger gate, calculate the product of the difference between the sensing gap degree and the gap degree threshold, the weight coefficient, the cross-scale stability, and the complement of the signal saliency, construct a correction compensation term for each of the sub-band coefficients, and superimpose the correction compensation term onto the signal saliency to obtain the calibrated saliency.

[0188] S4. Construct a mapping relationship between the perceived gap, the signal saliency and the calibration saliency, then modulate each sub-band coefficient and obtain the enhanced environmental image through inverse transformation.

[0189] This embodiment has the same beneficial effects as Embodiment 1.

[0190] Example 3

[0191] This embodiment introduces a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the low-light visual assistance enhancement method for glaucoma patients as described in Embodiment 2.

[0192] The low-light vision enhancement method for glaucoma patients described in Example 2 can be applied in software form, such as by designing it as a standalone program and installing it on a computer device, which can be a computer, a smart device, or the like. Alternatively, it can be designed as an embedded program and installed on a computer terminal, such as on a microcontroller.

[0193] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A low-light vision enhancement system for glaucoma patients, characterized in that, include: The data acquisition module is used to acquire environmental images, bias parameters for adjusting brightness compensation, and weight coefficients for controlling signal classification. The metric extraction module is used to perform multi-scale frequency band decomposition on the environmental image to obtain multiple sub-band coefficients; extract the average amplitude of the sub-band coefficients and the noise amplitude corresponding to each frequency band; calculate the perceptual gap based on the ratio of the average amplitude to the noise amplitude and the bias parameter; calculate the proportion of the square of the sub-band coefficient in the sum of the squares of the sub-band coefficients and the squares of the noise amplitude to obtain the signal saliency; calculate the sign consistency decision result between the sub-band coefficients of adjacent frequency bands, and the ratio of the minimum amplitude to the maximum amplitude between the sub-band coefficients, and then obtain the cross-scale stability. The decision coupling module is used to use the perceived gap degree being greater than a preset gap degree threshold as a trigger gate, calculate the product of the difference between the perceived gap degree and the gap degree threshold, the weight coefficient, the cross-scale stability, and the complement of the signal saliency, construct a correction compensation term for each of the sub-band coefficients, and superimpose the correction compensation term onto the signal saliency to obtain the calibrated saliency. The image reconstruction module is used to construct the mapping relationship between the perceived gap, the signal saliency and the calibration saliency, and then modulate each sub-band coefficient and obtain the enhanced environmental image through inverse transformation.

2. The low-light visual assistance enhancement system for glaucoma patients as described in claim 1, characterized in that, The calculation of the perceived gap degree in the metric extraction module specifically includes: The noise amplitude is linearly scaled using a preset clarity visibility coefficient to obtain a clarity energy threshold; Calculate the ratio of the average amplitude of the subband coefficient in the preset edge region to the clear energy threshold; Perform a negative correlation inversion mapping on the ratio, and obtain the scene component after taking the positive value; The scene components and the bias parameters are linearly offset and superimposed to obtain the composite notch. A nonlinear saturation constraint mapping is performed on the synthetic gap value to restrict the mapping result within a preset interval, thereby obtaining the perceived gap value.

3. The low-light visual assistance enhancement system for glaucoma patients as described in claim 1, characterized in that, The calculation of the sign consistency decision results between sub-band coefficients of adjacent frequency bands, and the ratio of the minimum amplitude to the maximum amplitude among the sub-band coefficients, to obtain cross-scale stability, specifically includes: Normalization of subband coefficients in different frequency bands is performed using a preset typical energy reference for the frequency band to obtain normalized directional coefficients; Extract the normalized directional coefficients of adjacent frequency bands at the same spatial location, and determine whether the signs of the normalized directional coefficients of the adjacent frequency bands are consistent. If they are consistent, output the first logical value; if they are inconsistent, output the second logical value. The result is the ratio between the minimum and maximum absolute values ​​of the normalized directional coefficients of the adjacent frequency bands. Multiply the first or second logic value with the ratio result and take the average value over all adjacent frequency bands to obtain the initial cross-scale stability. The initial cross-scale stability is compared with a preset stability threshold, and the final cross-scale stability is output based on the comparison result: When the initial cross-scale stability is higher than the preset stability threshold, the original value is output as the final cross-scale stability. When the initial cross-scale stability is not higher than the preset stability threshold, if the preset structural compensation triggering condition is met, the preset compensation value is output; otherwise, zero is output as the final cross-scale stability.

4. The low-light visual assistance enhancement system for glaucoma patients as described in claim 1, characterized in that, The decision coupling module specifically includes: When the sensing gap is greater than the gap threshold, the difference between the sensing gap and the gap threshold is extracted as the first modulation factor. Perform a reverse mapping on the signal saliency to obtain a second modulation factor; The first modulation factor is attenuated using the second modulation factor to obtain a first attenuation result; The cross-scale stability is extracted as a third modulation factor, and the first attenuation result is jointly multiplicatively modulated using the third modulation factor and weighting coefficients to obtain the correction compensation term. The correction compensation term is superimposed with the signal significance, and the numerical boundary truncation operation is performed on the superposition result to obtain the calibration significance.

5. The low-light visual assistance enhancement system for glaucoma patients as described in claim 1, characterized in that, The image reconstruction module constructs the mapping relationship to modulate the coefficients of each sub-band, specifically including: Asymmetric power calculation is performed on the perceived gap degree and the signal saliency to obtain the enhanced demand degree; Extract the complement of the calibration significance as the noise ratio, and calculate the mean of the sum of the preset value and the enhancement demand as the stress factor; The enhancement cost is obtained by multiplying the noise percentage with the pressure factor. The product of the enhancement requirement and the calibration significance is used as the gain numerator, and the sum of the gain numerator and the enhancement cost is used as the gain denominator; The initial gain weight is obtained by calculating the ratio of the gain numerator to the gain denominator, and the absolute value of the difference between the initial gain weights of adjacent frequency bands is calculated. When the absolute value of the difference is greater than a first preset threshold, propagation correction is triggered. The lower frequency band in adjacent frequency bands is defined as the low frequency band, and the higher frequency band is defined as the high frequency band; When the initial gain weight corresponding to the low frequency band is greater than the second preset threshold, if the signal saliency corresponding to the low frequency band is greater than the preset low frequency support threshold, then the calibration saliency corresponding to the low frequency band is extracted, and the absolute value of the difference, the calibration saliency and the preset propagation coefficient are multiplicatively modulated to obtain a positive gain correction amount, which is then superimposed on the initial gain weight of the high frequency band. When the initial gain weight corresponding to the high-frequency band is greater than the second preset threshold, if the signal salience corresponding to the low-frequency band is less than the preset false support criterion and the cross-scale stability of the high-frequency band is less than the preset high-frequency structure threshold, then the negative suppression factor is calculated using the signal salience of the low-frequency band, and the negative suppression factor is multiplied by the initial gain weight of the high-frequency band.

6. The low-light visual assistance enhancement system for glaucoma patients as described in claim 1, characterized in that, The system also includes a state storage unit and a feedback parameter update link. The state storage unit is used to store the bias parameter and the weight coefficient for the data acquisition module to read. The feedback parameter update chain includes: Acquire the gaze sequence and inertial motion characteristics output by smart wearable devices; The gaze point sequence is segmented based on a preset speed threshold to obtain gaze events, and the dwell time of the gaze events is extracted. Within a preset time interval after the gaze event ends, the rate of change of inertial motion features is extracted and judged according to preset judgment conditions to obtain the visual struggle state. A baseline duration is constructed and updated based on the dwell time and a preset baseline learning rate. When the dwell time of the gaze event is greater than the corresponding preset multiple of the baseline duration and the visual struggle state is met, a positive cognitive weight is assigned to the gaze event. Based on the normalized difference between the dwell time of each gaze event and the baseline duration, and combined with the cognitive weights corresponding to each gaze event, the corresponding feedback bias is calculated to update the bias parameters and weight coefficients in the state storage unit.

7. The low-light visual assistance enhancement system for glaucoma patients as described in claim 6, characterized in that, The feedback deviation is used to update the bias parameters in the state storage unit, specifically including: Extract the bias parameters from the previous update cycle; The bias parameter is subjected to a leakage integral attenuation process using a preset forgetting attenuation coefficient to obtain the attenuation reference component; The incremental driving component is obtained by multiplying the feedback deviation of the current cycle by the preset individual offset learning rate. The attenuation reference component and the incremental driving component are added together to synthesize the result, and a one-way non-negative interval truncation operation is performed on the synthesis result. The truncated result is used as the bias parameter of the current period for updating.

8. The low-light visual assistance enhancement system for glaucoma patients as described in claim 6, characterized in that, The feedback deviation is used to update the weight coefficients in the state storage unit, specifically including: Filter out the set of active cells whose number exceeds the preset activity threshold; Calculate the arithmetic mean of the feedback deviations corresponding to all grids within the active grid set to obtain the global feedback feature quantity; Calculate the deviation component between the weight coefficient of the previous update cycle and the preset weight coefficient benchmark value, and multiply the deviation component by the preset forgetting decay coefficient to obtain the steady-state regression force term. The weight coefficients are then superimposed on the product of the global feedback feature and the learning rate of the weight coefficients, and the steady-state regression force term is subtracted. After being truncated by boundary constraints, the weight coefficients are updated to the current period's weight coefficients.

9. A method for enhancing low-light vision for glaucoma patients, applied to the low-light vision enhancement system for glaucoma patients as described in any one of claims 1-8, characterized in that, Includes the following steps: Acquire environmental images, bias parameters for adjusting brightness compensation, and weight coefficients for controlling signal classification; The environmental image is decomposed into multiple sub-band coefficients at multiple scales. The average amplitude of the sub-band coefficients and the noise amplitude corresponding to each frequency band are extracted. The perceptual gap is calculated based on the ratio of the average amplitude to the noise amplitude and the bias parameter. The signal saliency is obtained by calculating the proportion of the square of the sub-band coefficient in the sum of the squares of the sub-band coefficients and the squares of the noise amplitude. The sign consistency decision results between the sub-band coefficients of adjacent frequency bands and the ratio of the minimum amplitude to the maximum amplitude between the sub-band coefficients are calculated to obtain the cross-scale stability. Using the sensing gap degree being greater than a preset gap degree threshold as a trigger gate, the product of the difference between the sensing gap degree and the gap degree threshold, the weight coefficient, the cross-scale stability, and the complement of the signal saliency is calculated to construct a correction compensation term for each of the sub-band coefficients, and the correction compensation term is superimposed on the signal saliency to obtain the calibrated saliency. A mapping relationship is constructed between the perceived gap, the signal saliency, and the calibration saliency. Then, the coefficients of each sub-band are modulated and subjected to inverse transformation to obtain the enhanced environmental image.

10. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the low-light visual assistance enhancement method for glaucoma patients as described in claim 9.

Images